Pritam Dalal
An ML algorithm learns complex patterns in high-dimensional space with out being specifically directed.
Classical Machine Learning: supervised and unsupervised
Deep Learning: neural networks and reinforcement learning
Deep-learning has many stunning success stories outside of finance (chat-bots, language translation, self-driving cars), but classical machine learning still dominates structured data.
A feed-forward neural network with a single hidden layer can approximate almost any function with arbitrary accuracy.
However you may need a lot of units in the hidden layer.
From a practical standpoint, you’re probably better off with multiple layers of modest size.
\[\begin{align*} \text{Forecast Error} \,\, = \,\, &\text{In-Sample Error} \, + \\[6pt] &\text{Model Instability} \, + \\[6pt] &\text{Random Error} \end{align*}\]
In-Sample Error: bias
Model Instability: variance
Random Error: irreducible noise
The art of machine learning is finding the right balance between bias and model stability.
Model flexibility comes at the cost of model instability.
Many of the ideas of machine learning have lived under the umbrella of statistics.
Substantive differences:
Substantive differences (cont’d):
Financial ML offers the ability to extract insights from data:
Warning: financial modeling is harder than driving cars or recognizing faces!
…machine learning is the most recent embodiment of the long standing quantitative investing paradigm - the idea of using data-driven approaches to build more efficient portfolios…it is a natural evolution of quantitative tools in asset management, and not a revolutionary shift in the business model.
Can Machines “Learn” Finance
Warning: returns prediction is harder than chess, go, and Atari.
Image recognition analogy: “…dogs morph into cats as soon as the model gets good at predicting dogs.”
High volatility due to unanticipated news.
Volatility is especially high relative to risk-premia which are small.
Any return predictability is wiped away by trading, and all that remains is noise.
Machine learning models can be hard to interpret, i.e. describe how an input observation results in a prediction.
Antidotes:
Choosing predictability vs interpretability of a model is ultimately a business decision.
The use of deep learning in asset management tends to be indirect:
Data providers like Refinitiv and Bloomberg have programmatic APIs have caused an explosion in data-driven finance:
It is useful to categorize alternative data by how it is generated:
Characteristics:
Concepts central to machine learning are largely absent from econometric literature: classifier, clustering, neural networks, machine learning
Correlation is the foundation of econometrics.
Here are the problems with correlations:
Most interesting data sets can be tackled by machine learning but not econometrics.
Econometrics models uninteresting data.
Traditional models don’t learn from the data.
Rather, we are required to provide the exact functional (parametric) specification in order to reach useful conclusions.
Feature importance in financial econometrics relies on \(p\)-values. However:
MDA is an ML technique for assessing feature importance that is more robust that \(p\)-values.
Using small amounts of structured data to make long-term forecasts.
Forecasts rely on statistical relationships between lagged observation and future outcome.
Using a large amount of unstructured data to make short-term predictions:
MLOps is a set of practices that aims to deploy and maintain machine learning models in production reliably and efficiently.
Wikipedia
Somewhere a quant has accomplished the following:
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
X = pd.read_csv('features.csv')
y = pd.read_csv('labels.csv')
model = RandomForestClassifier()
model.fit(X, y)You probably want to:
RStudio is becoming Posit!
This presentation was created with Quarto in a Jupyter Notebook. No R!
Over a decade of Shiny development brought into the Python ecosystem.
Advances in Financial Machine Learning, Marcos López de Prado (2018)
Artificial Intelligence in Finance, Yves Hilpsich (2020)
Big Data and AI Strategies, JP Morgan (2017)
Can Machines “Learn” Finance?, Ronen Israel, Bryan Kelly, Tobias Moskowitz (2020)